Study On Algorithms Of Mixed Spectral Simulation And Unmixing For Hyperspectral Remote Sensing Data

Due to the complexity and diversity of ground objects and sensor’s limitation of spatial resolution, mixed pixels are widely exist in remote sensing data and thus influencing the applICAtion accuracy on analyzing ground target and extracting information. The most effective way to solve mixed pixel problem is mixed pixel decomposition. It is implemented by spectral mixture models based on particular physICAl or statistICAl theories. Spectral mixture models simulate the formation process of mixed pixel and break through the spatial resolution of hyperspectral image to obtain the endmember spectra and corresponding abundance fractions, aiming to improvethe accuracy of analysis on ground object and hyperspectral image classifICAtion. From the perspective of formation and unmixing of mixed spectra, the spectral model can be divided into two different types: spectral mixing model and spectral unmixing model. Spectral mixing model is a forward modeling, which describles the process of light transmitting and eventually generating the reflectance in the sensor. While spectral unmixing model is the inversion of spectral mixing model. The two models are interactive and inseparable. In the view of light transporting principle, spectral mixture models are divided into two different parts, linear spectral model(LSM) and nonlinear spectral model(NLSM). LSM is simple and has clear physICAl signifICAnce by simplifying the light transfer process. In order to obtain the information of a pixel, it assumes that the mixed spectra of a pixel is linear combination of endmember spectra and its corresponding abundance. NLSM takes more factors into account and simulates the transfer process as accurate as possible so that to generate more accurate results. The two different models have its own advantages. Based on the researches of the two models, we propose several different approaches for hyperspectral unmixing by analyzing spatial features and spectral features of remote sensing data.By considering the effects of the endmember’s spatial location on mixed spectra, the weight coefficients of different area are incorporated in a new proposed method, equidistant/homalographic model, for mixed spectral simulation. To validate the proposed model, the experiments based on the reflex platform and FieldSpec 3 Hi-Res portable spectrum instrument are implemented by fixing the target area and geometrICAl observation conditions to detect the mixed spectra of calcite and green leaves. We have conducted two different types of experiments: regular shape leaf mixed with calcite and nonregular shape leaf mixed with calcite. The results from analysis of measured mixed spectra show that the closer the distance between probe and detected point, the higher the weight coefficient of contribution to mixed spectra is.By equidistant/homalographic experiment, the weight coefficients of constribution of different detected area are calculated by mixed spectra of calcite and black board for mixed spectral simulation. By comparson, the equidistant/homalographic model gengerates more accuracy results based on the effects of spatial feature and also provides a new method for hyperspectral unmixing and a new approach to eliminate the limitations that are aroused by topographic inequality.Different materials have different characteristics at different bands, which can be used to identify ground objects by analyzing the differrence of spectral reflectance. The absorption feature of calcite is located aroud 2300 nm while the green vegetation is a reflective peak around those bands. When the cover proportion of green vegetation increases, the spectral absorption feature will decrease in the mixed spectra. Under the purpose of eliminating the effects on information extraction accuracy, the improved linear spectral mixture model is proposed to improve the simulation accuracy by considering leaf transmissivity based on measured data. To prove the improved model is suitable for spectral analysis, two different experiments are conducted to simulate the mixed spectra. One is the transmission-free leaf mix with calcite and the other one is transmission leaf mixing with calcite. Through analysis, the rule of absorption feature change with the cover proportion of leaf, which is obtained by processing the measured mixed spectra that mixed by endmember spectra from USGS library, is applied to Hyperion image data for inversion of carbonate minerals. The inversion results are validated by geologic map and sample point in the study area. It indICAtes that the absorption feature extract specified minerals successfully and provide a novel approach for inversion of minerals with distinguished features.Blind source separation(BSS) algorithm is a method that separate the mixed signals under an unknown situration of endmember spectra and its abundance fractions. Nonnegative matrix factorization(NMF) is one of the most widely used BSS algorithms. Based on LSM and nonnegativity of hyperspectral data, NMF method decomposes the mixed spectral matrix into a multiplICAtion of spectral matrix and abundance matrix. Due to the nonconvexity of NMF, the separated result is not unique. In order to solve this problem and improve the accuracy and algorithm efficiency, an approach that is called constrained multilayer NMF(CMLNMF) is proposed for hyperspectral unmixing. CMLNMF takes minimum volume constraint on spectral matrix and sparseness constraint on abundance matrix and also take hierarchICAl processing on abundance matrix which make the algorithm more meaningful and more efficiency. By comparing the outcomes from simulated data and real image data with other four different NMF methods, CMLNMF has a better performance and also has a relative higher effective process.Independent component analysis(ICA) is other widely used method among BSS algorithms. ICA assumes that all the source signals are statiscally independent and its purpose is to get the source signals as independent as possible by using iterative process. It has been widely used in the aera of signal separation and image processing. But tradional ICA is not suitable for remote sensing data processing because of the negative value and amplitude ambiguity by mean value and whitening processing. To improve the efficiency and accuracy of ICA, FastICA is used for hyperspectral unmixing based on the purpose of identifying target on the sea and operation of regarding the endmember spectra as independent source signal. For comparison, we take analysis on the impact factors that may influence the separated reults based on FastICA method. Through changing differnet parameters, the error analysis of separated results is evaluated by coefficient of determination. The outcomes show that initial matrix, the shape of spectra and signal-to-noise ratio have signifICAnt impacts on separated results and thus providing several thoughts for improving ICA and a new method for target identifICAtion on the sea.Based on the feature of hyperspectral data, nonnegativity and sum-to-one constraints were proposed as two constraints for spectral unmixing model. Generally, the two constraints are expressed by some specified functions so that the iterative calculation makes all the constraints work together. It requires the functions accurate enough. Otherwise, the separated results will divergence or incorrect. To solve the problem, GeometrICAl constrained ICA(GCICA) is proposed from the view of geometrICAl points. GCICA separates the processing of two constraints. It takes normalization processing to make the abundance sum-to-one first. This step can restrain the sum-to-one constraint thoroughly. Then, the correct results are generated by constraint of nonnegativity and minimum of mutual information. Furthermore, the initialization of spectral matrix is processed for GCICA based on the analysis of impact factors to FastICA, contributing to improve the efficiency and accuracy of the proposed algorithm. Through validation of simulated data and real image data, GCICA performs better in endmember spectra and abundance extraction and also offer an effective way for hyperspectral unmixing.